Hi guys!
I have a question. So I am trying to implement SIFT.
My question is regarding inter-pixel distance. I think after downsampling the number of pixels change but the inter-pixel distance remains the same. But I am reading the paper 'Anatomy of SIFT' which states that after downsampling(subsampling) by 2 the inter-pixel distance doubles.
Am I missing something?
Thanks!
Hi,
I think the problem may be here that you are using different definitions of the concept inter-pixel distance.
Inter-pixel distance is not a standard concept at all. To give you an idea in a camera sensor the inter-pixel distance is the distance in mm from the center of a photo detector to the next. As camera sensors have standard size, the inter-pixel distance is inversely proportional to the resolution.
The same if you talk about a screen representation of a digital image. Where it effectively doubles when you subsample an image.
Of course, in a different context it may be the opposite, if you are talking, for example of texture features, after upsampling an image, two pixels that were located at a distance 1 from each other will be now located at a distance 2 and in other situations argue about inter-pixel distance in relation with subsampling, may have no sense at all.
After taking a brief look to the paper you mention. I would say that unfortunately, such a specific and detailed paper falls in the trend of many scientific documents in several places. That is, it was not written to be easy to read or even understandable to outsiders, as a lot of the concepts they use are never explained.
Hi,
I think the problem may be here that you are using different definitions of the concept inter-pixel distance.
Inter-pixel distance is not a standard concept at all. To give you an idea in a camera sensor the inter-pixel distance is the distance in mm from the center of a photo detector to the next. As camera sensors have standard size, the inter-pixel distance is inversely proportional to the resolution.
The same if you talk about a screen representation of a digital image. Where it effectively doubles when you subsample an image.
Of course, in a different context it may be the opposite, if you are talking, for example of texture features, after upsampling an image, two pixels that were located at a distance 1 from each other will be now located at a distance 2 and in other situations argue about inter-pixel distance in relation with subsampling, may have no sense at all.
After taking a brief look to the paper you mention. I would say that unfortunately, such a specific and detailed paper falls in the trend of many scientific documents in several places. That is, it was not written to be easy to read or even understandable to outsiders, as a lot of the concepts they use are never explained.
Dear Mr. Ghost,
I guess it's about the following.
With 2-times downsampling, two adjacent pixels will give the image of two pixels that were initially twice the distance. It is clear that
downsampling leads to a data loss.
For clarity I recommend the article:
https://www.researchgate.net/publication/235861975_Analysis_of_quantitative_properties_of_ultrasound_images
Best regards
K.Yankov
I think that the point is that SIFT descriptors and local keypoints are extracted by analysing local maxima and minima in DoG (Difference of Gaussians) functions on different scales of the image, that is, the down-sampled version of the image itself. So the distance between two pixel at the first scale is then down-sampled as well as the pixels.
- Badges
- Science topic
- Similar topics
- Engineering
- Electrical Engineering
- Signal Processing
- Image Processing